Short answer
Smart-meter phase identification investigates which supply phase is associated with a customer record. It is different from allocating demand to a transformer or deciding which customers should change phase. Transformer-side reference data can help resolve incorrect labels, but a correlation result remains evidence to validate. Reliable identification requires known reference labels, compatible measurements and a clear treatment of ambiguous records.
Define the data relationship first
Confirm the customer-to-transformer association and the phase labels of the reference channels. For a paired voltage-correlation calculation, align customer and reference timestamps, interval lengths and missing-data treatment within that measurement campaign. A billing-period energy total is not interchangeable with a sampled voltage series.
Customer voltage and transformer-terminal voltage are measured at different locations. Branch loading, PV, neutral movement and common upstream changes can alter their relationship. Correlation compares patterns; it does not assert that the two voltage magnitudes should be equal. Phase mapping from GIS supplies a starting hypothesis, not independent confirmation.
What the original research supports
Hoogsteyn and co-authors compared voltage- and power-based methods, reference choices and ensemble approaches under specified feeder and meter scenarios. Their numerical results demonstrate that data quality and the selected reference matter. They do not establish one accuracy figure for another network.
Some studied ensembles combine voltage and power campaigns from different periods. That is different from correlating two unaligned voltage traces sample by sample. Keep the method’s input requirements explicit rather than applying one synchronization rule indiscriminately to every algorithm.
An independent arithmetic illustration
Assumptions. Four synthetic voltage-variation samples are dimensionless and aligned. They are not customer records. Reference A = [1, −1, 1, −1], B = [1, 1, −1, −1] and C = [1, −1, −1, 1]. The customer pattern is [0.8, −0.9, 1.1, −1.0].
Use normalized Pearson correlation: subtract each series mean, sum the paired products, and divide by the square root of the product of their centered sums of squares. The resulting score lies between −1 and +1 for nonconstant series.
| Reference | Correlation |
|---|---|
| A | About +0.993 |
| B | About −0.052 |
| C | About −0.105 |
A matches this invented pattern best. Four samples cannot establish confidence, field accuracy or a defensible assignment threshold. The calculation is independent of the paper’s numerical experiments and does not reproduce its algorithm settings.
Preserve uncertain cases
Common voltage movement can make several phase references correlate strongly. Weak variation, clock errors, estimated readings, branch changes and incomplete meter coverage can also make the ranking unstable. Constant series have undefined correlation and should not be assigned from this calculation. Retain an unresolved category instead of forcing every customer into A, B or C.
Validate results against independently known labels and separate periods, with a documented rule for conflicts and mapping revisions. Protect customer data and record the evidence behind each accepted change. Phase swapping is a later engineering and authorized field-work decision; a data-based phase label does not grant permission to alter an energized installation.
Sources and scope
- A. Hoogsteyn, M. Vanin, A. Koirala and D. Van Hertem, Low Voltage Customer Phase Identification Methods Based on Smart Meter Data, original author preprint, arXiv:2204.06372v1, 2022.